Infrastructure & orchestration

Intelligent inference routing and capacity orchestration.

Ariadne Router and Ariadne ModelHost work together to orchestrate your AI infrastructure. Direct requests dynamically, provision model endpoints on demand, and seamlessly span your dedicated hardware and the cloud.

How it works

The engine behind the agents.

Whether you are running your own hardware clusters or relying on managed endpoints, Ariadne's infrastructure layer ensures that every prompt reaches the right model at the right time.

Ariadne Router

An interactive inference gateway that directs model requests to the best available host. It supports caching affinity, priority weights, and least-used endpoint selection to balance load intelligently.

It features native support for OpenAI and Anthropic APIs, plus specialist integrations for Azure AI Foundry, AWS Bedrock, and Alibaba Dashscope. Route seamlessly to any compatible SaaS provider like OpenRouter, while orchestrating Ariadne ModelHost to provision your local inference capacity.

  • Intelligent routing: Balances traffic across multiple identical models.
  • Access control: Protects endpoints with robust API-key authentication.
  • Admin dashboard: Manage your host registry and view live traffic through a dedicated management interface.
  • Observability: Capture and replay traffic for auditing or debugging.

Ariadne ModelHost

A lightweight model-management service that acts as the runtime environment for your hardware. It receives instructions from the Router to provision and launch new model instances dynamically.

  • Dynamic provisioning: Automatically downloads models from registries and sets them up.
  • Runtime recipes: Launches and manages underlying inference servers on demand.
  • Hardware isolation: Spins up instances on isolated ports for secure, multi-tenant scaling.
  • API controlled: Fully managed via a secure HTTP control plane.

Supported runtimes

ModelHost includes native recipe support for TensorSharp out of the box, handling both high-performance GPU execution and inexpensive CPU/GGML fallbacks.

It also seamlessly orchestrates containerised inference engines including vLLM, SGLang, Ollama, and NVIDIA NIM via Docker.

Because the recipe system is fully extensible, it can be configured to wrap almost any standard HTTP-based local inference service.

Inference delegation

Orchestrating, not executing.

Ariadne ModelHost is a lifecycle manager, not a native inference engine.

Rather than reinventing model execution, ModelHost delegates the heavy lifting to dedicated, high-performance inference servers. It uses flexible runtime recipes to automatically download models from Hugging Face and boot the underlying engine on an isolated, securely allocated port.

This clean boundary ensures that ModelHost remains lightweight and stable, focusing entirely on provisioning, health monitoring, and scaling—while letting specialised engines handle the actual token generation.

Deployment strategy

Orchestration & hybrid scaling

Build a resilient, cost-effective infrastructure that automatically adapts to your workload.

The combination of Ariadne Router and ModelHost allows you to establish strict prioritisation rules. Keep your critical and predictable workloads running on your reserved, on-premise hardware to maximise ROI.

When demand spikes, the Router automatically shifts excess load, instructing ModelHost to spin up new local instances or seamlessly failing over to managed cloud providers. Health checks and route selection happen transparently—ensuring uptime without manual intervention.

Prioritise local capacity

Saturate your fixed-cost, on-premise GPUs first before spending on the cloud.

Burst to cloud

Automatically scale out to cloud APIs when local queues build up during peak demand.

CPU fallbacks

Leverage inexpensive CPU instances locally for background or low-priority tasks.